TensorTrait

功能说明

GlobalTensor和LocalTensor中通过ShapeInfo类型的成员变量来保存shape信息，可以通过SetShapeInfo、GetShapeInfo来进行设置或者获取，通常用于算子实现内部的shape信息保存和传递。在不使用上述ShapeInfo功能的情况下，不需要这些信息。此时可以使用TensorTrait定义不含ShapeInfo的GlobalTensor以及LocalTensor，以降低内存占用，提升运行性能。

定义原型

template <typename T>
struct TensorTrait {
    using LiteType = T;
};

参数说明

表1 TensorTrait结构体模板参数说明
参数名	描述
T	只支持如下基础数据类型：int4b_t、uint8_t、int8_t、int16_t、uint16_t、bfloat16_t、int32_t、uint32_t、int64_t、uint64_t、float、half 。通过TensorTrait可以得到一个使用TensorTrait表达的Tensor数据类型：在TensorTrait结构体内部，使用using关键字定义了一个类型别名LiteType，与模板参数T类型一致。通过TensorTrait定义的LocalTensor/GlobalTensor不包含ShapeInfo信息。例如： LocalTensor<float>对应的不含ShapeInfo信息的Tensor为LocalTensor<TensorTrait<float>>。

约束说明

同一接口不支持同时输入TensorTrait类型的GlobalTensor/LocalTensor和非TensorTrait类型的GlobalTensor/LocalTensor。
非TensorTrait类型和TensorTrait类型的GlobalTensor/LocalTensor相互之间不支持拷贝构造和赋值运算符。

TensorTrait特性当前仅支持如下接口：

表2 TensorTrait特性支持的接口列表
接口分类	接口名称	备注
基础API>内存管理与同步控制>TQue/TQueBind	AllocTensor、FreeTensor、EnQue、DeQue	_
基础API>矢量计算>单目指令	Exp、Ln、Abs、Reciprocal、Sqrt、Rsqrt、Not、Relu	-
基础API>矢量计算>双目指令	Add、Sub、Mul、Div、Max、Min、And、Or、AddRelu、AddReluCast、AddDeqRelu、SubRelu、SubReluCast、MulAddDst、FusedMulAdd、FusedMulAddRelu、	-
基础API>矢量计算>标量双目指令	Adds、Muls、Maxs、Mins、ShiftLeft、ShiftRight、LeakyRelu	-
基础API>数据搬运	DataCopy、Copy	切片数据搬运接口需要ShapeInfo信息，不支持输入TensorTrait类型的GlobalTensor/LocalTensor
基础指令>ISASI（体系结构相关）>矩阵计算	InitConstValue、LoadData、LoadDataWithTranspose、SetAippFunctions、LoadImageToLocal、LoadUnzipIndex、LoadDataUnzip、LoadDataWithSparse、SetFmatrix、SetLoadDataBoundary、SetLoadDataRepeat、SetLoadDataPaddingValue、Mmad、MmadWithSparse、Fixpipe、SetFixPipeConfig、SetFixpipeNz2ndFlag、SetFixpipePreQuantFlag、BroadCastVecToMM、SetHF32Mode、SetHF32TransMode、SetMMLayoutTransform、CheckLocalMemoryIA、Conv2D、Gemm	-

调用示例

双目指令使用TensorTrait样例

// 使用系统描述符TensorTrait
AscendC::LocalTensor<AscendC::TensorTrait<half>> tensor1 = que1.DeQue<AscendC::TensorTrait<half>>();
AscendC::LocalTensor<AscendC::TensorTrait<half>> tensor2 = que2.DeQue<AscendC::TensorTrait<half>>();
AscendC::LocalTensor<AscendC::TensorTrait<half>> tensor3 = que3.AllocTensor<AscendC::TensorTrait<half>>();
Add(tensor3, tensor1, tensor2, tensor3.GetSize());

标量双目指令使用TensorTrait样例

#include "kernel_operator.h"
class KernelBinaryScalarTrait {
public:
    __aicore__ inline KernelBinaryScalarTrait() {}
    __aicore__ inline void Init(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
    {
        srcGlobal.SetGlobalBuffer((__gm__ int16_t*)src);
        dstGlobal.SetGlobalBuffer((__gm__ int16_t*)dstGm);
        pipe.InitBuffer(inQueueSrc, 1, 512 * sizeof(int16_t));
        pipe.InitBuffer(outQueueDst, 1, 512 * sizeof(int16_t));
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }
private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<int16_t>> srcLocal = inQueueSrc.AllocTensor<AscendC::TensorTrait<int16_t>>();
        AscendC::DataCopy(srcLocal, srcGlobal, 512);
        inQueueSrc.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<int16_t>> srcLocal = inQueueSrc.DeQue<AscendC::TensorTrait<int16_t>>();
        AscendC::LocalTensor<AscendC::TensorTrait<int16_t>> dstLocal = outQueueDst.AllocTensor<AscendC::TensorTrait<int16_t>>();

        uint64_t mask = 128;
        int16_t scalar = 2;
        // repeatTimes = 4, 128 elements one repeat, 512 elements total
       // dstBlkStride, srcBlkStride = 1, no gap between blocks in one repeat
       // dstRepStride, srcRepStride =8, no gap between repeats
        AscendC::Adds(dstLocal, srcLocal, scalar, mask, 4, {1, 1, 8, 8});
        
        outQueueDst.EnQue(dstLocal);
        inQueueSrc.FreeTensor(srcLocal);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<int16_t>> dstLocal = outQueueDst.DeQue<AscendC::TensorTrait<int16_t>>();
        AscendC::DataCopy(dstGlobal, dstLocal, 512);
        outQueueDst.FreeTensor(dstLocal);
    }
private:
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::QuePosition::VECIN, 1> inQueueSrc;
    AscendC::TQue<AscendC::QuePosition::VECOUT, 1> outQueueDst;
    AscendC::GlobalTensor<AscendC::TensorTrait<int16_t>> srcGlobal, dstGlobal;
};
extern "C" __global__ __aicore__ void binary_scalar_trait_kernel(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
{
    KernelBinaryScalarTrait op;
    op.Init(src, dstGm);
    op.Process();
}

矩阵计算基础API使用TensorTrait样例

#include "kernel_operator.h"
template <typename dst_T, typename fmap_T, typename weight_T, typename dstCO1_T, typename bias_T> class KernelMatmul {
public:
    __aicore__ inline KernelMatmul(uint16_t mIn, uint8_t kIn, uint8_t nIn, bool initl1In, bool initl0In)
    {
        m = mIn;
        k = kIn;
        n = nIn;
        aSize = m * k;
        bSize = k * n;
        cSize = m * n;
        initl0 = initl0In;
        initl1 = initl1In;
    }
    __aicore__ inline void Init(__gm__ uint8_t *a, __gm__ uint8_t *b, __gm__ uint8_t *c)
    {
        aGM.SetGlobalBuffer((__gm__ fmap_T *)a);
        bGM.SetGlobalBuffer((__gm__ weight_T *)b);
        cGM.SetGlobalBuffer((__gm__ dstCO1_T *)c);
        pipe.InitBuffer(inQueueA1, 1, aSize * sizeof(fmap_T));
        pipe.InitBuffer(inQueueA2, 1, aSize * sizeof(fmap_T));
        pipe.InitBuffer(inQueueB1, 1, bSize * sizeof(weight_T));
        pipe.InitBuffer(inQueueB2, 2, bSize * sizeof(weight_T));
        pipe.InitBuffer(outQueueCO1, 1, cSize * sizeof(dstCO1_T));
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        SplitA();
        SplitB();
        Compute();
        CopyOut();
    }
private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<fmap_T>> a1Local = inQueueA1.AllocTensor<AscendC::TensorTrait<fmap_T>>();
        AscendC::LocalTensor<AscendC::TensorTrait<weight_T>> b1Local = inQueueB1.AllocTensor<AscendC::TensorTrait<weight_T>>();
        if(initl1 == true) {
            AscendC::InitConstValue(a1Local, {static_cast<uint16_t>(m * k * sizeof(fmap_T) / 32), 1, 0, 1});
            AscendC::InitConstValue(b1Local, {static_cast<uint16_t>(k * n * sizeof(weight_T) / 32), 1, 0, 1});
        } else {
            AscendC::DataCopy(a1Local, aGM, aSize);
            AscendC::DataCopy(b1Local, bGM, bSize);
        }
        inQueueA1.EnQue(a1Local);
        inQueueB1.EnQue(b1Local);
    }
    __aicore__ inline void SplitA()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<fmap_T>> a1Local = inQueueA1.DeQue<AscendC::TensorTrait<fmap_T>>();
        AscendC::LocalTensor<AscendC::TensorTrait<fmap_T>> a2Local = inQueueA2.AllocTensor<AscendC::TensorTrait<fmap_T>>();
        // 1、load2d L1->L0A
        AscendC::LoadData2dParams loadL0AParams;
        loadL0AParams.repeatTimes = m * k * sizeof(fmap_T) / 512;
        loadL0AParams.srcStride = 1;
        loadL0AParams.dstGap = 0;
        if (initl0 == true) {
            InitConstValue(a2Local, {static_cast<uint16_t>(m * k * sizeof(fmap_T) / 512), 1, 0, 1});
        } else{
            LoadData(a2Local, a1Local, loadL0AParams);
        }
        inQueueA2.EnQue<AscendC::TensorTrait<fmap_T>>(a2Local);
        inQueueA1.FreeTensor(a1Local);
    }
    __aicore__ inline void SplitB()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<weight_T>> b1Local = inQueueB1.DeQue<AscendC::TensorTrait<weight_T>>();
        AscendC::LocalTensor<AscendC::TensorTrait<weight_T>> b2Local = inQueueB2.AllocTensor<AscendC::TensorTrait<weight_T>>();
        // 2、load2d L1->L0B
        AscendC::LoadData2dParams loadL0BParams;
        loadL0BParams.repeatTimes = k * n * sizeof(weight_T) / 512;
        loadL0BParams.srcStride = 1;
        loadL0BParams.dstGap = 0;
        if (initl0 == true) {
            AscendC::InitConstValue(b2Local, {static_cast<uint16_t>(k * n * sizeof(weight_T) / 512), 1, 0, 1});
        } else{
            AscendC::LoadData(b2Local, b1Local, loadL0BParams);
        }
        inQueueB1.FreeTensor(b1Local);
        inQueueB2.EnQue<AscendC::TensorTrait<weight_T>>(b2Local);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<fmap_T>> a2Local = inQueueA2.DeQue<AscendC::TensorTrait<fmap_T>>();
        AscendC::LocalTensor<AscendC::TensorTrait<weight_T>> b2Local = inQueueB2.DeQue<AscendC::TensorTrait<weight_T>>();
        AscendC::LocalTensor<AscendC::TensorTrait<dstCO1_T>> c1Local = outQueueCO1.AllocTensor<AscendC::TensorTrait<dstCO1_T>>();
        mmadParams.isBias = false;
        mmadParams.m = m;
        mmadParams.n = n;
        mmadParams.k = k;
        AscendC::Mmad(c1Local, a2Local, b2Local, mmadParams); // m*n
        outQueueCO1.EnQue<AscendC::TensorTrait<dstCO1_T>>(c1Local);
        inQueueA2.FreeTensor(a2Local);
        inQueueB2.FreeTensor(b2Local);
    }
#if __CCE_AICORE__ <= 200
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<dstCO1_T>> c1Local = outQueueCO1.DeQue<AscendC::TensorTrait<dstCO1_T>>();
        uint16_t M_ = Ceil(m, 16) * 16;
        AscendC::LocalTensor<AscendC::TensorTrait<dst_T>> ublocal;
        AscendC::TBuffAddr tbufublocal;
        tbufublocal.logicPos = (uint8_t)AscendC::QuePosition::C1;
        ublocal.SetAddr(tbufublocal);
        ublocal.InitBuffer(0, M_ * n);
        DataCopyParams dataCopyParams;
        dataCopyParams.blockCount = 1;
        dataCopyParams.blockLen = Ceil(M_ * n * 4, 1024);
        DataCopyEnhancedParams enhancedParams;
        enhancedParams.blockMode = AscendC::BlockMode::BLOCK_MODE_MATRIX;
        AscendC::DataCopy(ublocal, c1Local, dataCopyParams, enhancedParams);
        PipeBarrier<PIPE_ALL>();
        outQueueCO1.FreeTensor(c1Local);
        dataCopyParams.blockCount = 1;
        dataCopyParams.blockLen = m * n *sizeof(dstCO1_T) / ONE_BLK_SIZE;
        dataCopyParams.srcStride = 0;
        dataCopyParams.dstStride = 0;
        AscendC::DataCopy(cGM, ublocal, dataCopyParams);
    }
#else
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<AscendC::TensorTrait<dstCO1_T>> c1Local = outQueueCO1.DeQue<AscendC::TensorTrait<dstCO1_T>>();
        AscendC::FixpipeParamsV220 fixpipeParams;
        fixpipeParams.nSize = n;
        fixpipeParams.mSize = m;
        fixpipeParams.srcStride = m;
        fixpipeParams.dstStride = n;
        fixpipeParams.ndNum = 1;
        fixpipeParams.srcNdStride = 0;
        fixpipeParams.dstNdStride = 0;
        AscendC::Fixpipe(cGM, c1Local, fixpipeParams);
        outQueueCO1.FreeTensor(c1Local);
    }
#endif
private:
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::QuePosition::A1, 1> inQueueA1;
    AscendC::TQue<AscendC::QuePosition::A2, 1> inQueueA2;
    AscendC::TQue<AscendC::QuePosition::B1, 1> inQueueB1;
    AscendC::TQue<AscendC::QuePosition::B2, 1> inQueueB2;
    // dst queue
    AscendC::TQue<AscendC::QuePosition::CO1, 1> outQueueCO1;
    AscendC::GlobalTensor<AscendC::TensorTrait<fmap_T>> aGM;
    AscendC::GlobalTensor<AscendC::TensorTrait<weight_T>> bGM;
    AscendC::GlobalTensor<AscendC::TensorTrait<dst_T>> cGM;
    uint16_t m, k, n;
    bool initl0, initl1;
    uint16_t aSize, bSize, cSize, b2Size;
    AscendC::MmadParams mmadParams;
};
extern "C" __global__ __aicore__ void cube_initconstvalue_simple_operator_half_16_32_16_true_false(
    __gm__ uint8_t *a, __gm__ uint8_t *b, __gm__ uint8_t *c)
{
    if ASCEND_IS_AIV {
        return;
    }
    KernelMatmul<float, half, half, float, half> op(16, 32, 16, true, false);
    op.Init(a, b, c);
    op.Process();
}

父主题： 数据类型定义